Kimi K2.7 Code on EU servers: what the model does and how lowcloud runs it
Kimi K2.7 Code is a large language model from the Chinese vendor Moonshot AI, published in June 2026. The vendor built it for two jobs: writing and changing program code, and calling tools, meaning it drives other programs and interfaces and works on with what they return. The weights, the trained numeric values that make up the model, are public on Hugging Face. The context window holds 256,000 tokens. The context window is how much text the model can take into account in one run; a token is the smallest unit it works with, roughly a short word or part of a word.
At lowcloud you pick Kimi K2.7 Code as the model behind an agent, for a single chat message, or for a single step of a workflow. Requests then go to the European provider Melious instead of to Moonshot AI, and are processed on servers inside the EU.
- Vendor
- Moonshot AI, China
- Context window
- 256k Token
- Reads images
- yes
- Licence
- Modified MIT
- Processing
- EU servers via Melious
Where the model comes from
Moonshot AI published the weights on Hugging Face in June 2026. The repository there was created on 11 June 2026 and last changed on 15 June 2026. There is no separate blog post for this model, so an exact announcement date is not documented. On the build: Kimi K2.7 Code is a mixture-of-experts model, MoE for short. The model is made of many sub-networks, and only a small share of them does the work for any given token. Moonshot lists 384 such sub-networks with 8 used per token, around one trillion parameters in total and 32 billion of them active per token. For you this means one thing: the model is very large, but answers faster than its total size suggests. The license is not MIT, even though many overview sites say so. Moonshot calls it a modified MIT license. The text matches MIT except for one addition: anyone using the model in a commercial product with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue has to show the name Kimi K2.7 Code in the user interface. Neither threshold has any practical effect for your work or for lowcloud. Calling it open source without qualification is still wrong. There is a newer model from the same vendor: Kimi K3, announced on 16 July 2026, with roughly 1 million tokens of context. That does not retire Kimi K2.7 Code. Moonshot still lists it in its own documentation as the model for code and tool work, and both are in the lowcloud EU catalog.
What "on EU servers" means here
The request does not go to Moonshot AI. The chain is: your client, then lowcloud UG in Bielefeld, then Melious AI GmbH in Saarbrücken, then a data centre inside the EU. Melious is our provider for every EU-hosted model. Melious runs no data centres of its own. It distributes to European operators and commits to processing staying inside the EU. What we do not know: which of those operators serves Kimi K2.7 Code. Melious does not publish it. We also do not know which exact build of the weights runs there. We pass on the commitment we have, and nothing beyond it. Before you use Kimi K2.7 Code with a client's data, we settle the contractual side with you: data processing under Art. 28 GDPR and the naming of sub-processors. That belongs in a conversation, not on a web page.
What Kimi K2.7 Code is good for
The vendor calls it a coding model. You do not have to write code to benefit. The two abilities it was trained for show up in workflows all the time. First, moving data from one shape into another. Splitting an invoice into separate fields. Reordering a table. Turning "14 March" into "2026-03-14". Turning a long email into a list of points. That follows the same strict rules as program code, and this is exactly what the model was tuned for. It sticks closer to format instructions in your prompt than a general-purpose model does. Second, calling tools. An agent that queries a CRM, checks the result, then runs a search and only then writes an email is calling several interfaces one after another. Moonshot built the model for chains like that: several calls in sequence, with a thinking step in between. It also reads images. Screenshots, scans and photos can be evaluated. We tested that with images, not with video. Moonshot publishes its own measurements for code and agent tasks on the model page. Those are vendor numbers, undated and not independently checked. We do not repeat them here.
The thinking step cannot be switched off
Kimi K2.7 Code always thinks. Moonshot has the thinking mode permanently enabled, and the interface offers no way to turn it off. Before every answer the model first produces a chain of reasoning, then the result. That has two consequences for your agent. Each step takes longer than with a model that has no thinking mode. And it consumes more tokens, because the reasoning counts too. On short, repetitive steps such as "rewrite these three fields", you pay for that without getting anything back. A second point from Moonshot's documentation: you cannot force the model to use a tool. It decides for itself whether to call one. If a step must trigger a lookup, build that into the workflow rather than relying on the model.
What it is not good for
Fixed JSON schemas. A JSON schema is a template that tells the model which fields its answer must contain. The vendor rejects requests carrying such a template with an error. lowcloud tested this against the interface itself; it is not stated on the vendor's page. Moonshot documents only the simpler JSON mode: the answer has to be valid JSON, but a fixed field list cannot be enforced. In that case lowcloud asks for the structure in the prompt instead. For steps whose output must match a specification exactly, pick a model that supports schemas: Kimi K2.6, GLM 5.1, GLM 5, Qwen 3.5 397B or DeepSeek V4 Flash, all of them also in the EU. Very large inputs. 256,000 tokens is a lot, but not enough to hand over a whole document archive in one run. If you need more, take Kimi K3, GLM 5.3, GLM 5.2 or DeepSeek V4 Pro. Those four hold roughly 1 million tokens. Large numbers of short, routine steps. Because of the permanent thinking mode, the model is slower and more expensive than necessary for those. Whether Kimi K2.7 Code writes weaker free-form client text than a general-purpose model is something we have not measured. The vendor positions it explicitly for code and tool use. If a step is mainly about wording, compare it in the chat against a general model from the catalog.
What it costs
There is no price on this page. What a single run cost is shown in euros next to every run in the app. That is the number you bill your client from. Within the EU catalog, Kimi K2.7 Code sits in the lower middle. Cheaper options include DeepSeek V4 Flash, which lowcloud uses in workspaces where no model has been picked. Kimi K3 is more expensive. Factor in the permanent thinking mode. Because every call produces extra tokens for the reasoning, a run can cost more than its position in the catalog suggests. The comparison that counts is between two real runs in the app, not between two places in a list.
Try Kimi K2.7 Code
Picking a model needs a free account. You then choose Kimi K2.7 Code in the chat, per agent, or per step.
Common questions
What is Kimi K2.7 Code?
A large language model from Moonshot AI in China, published in June 2026. It is built for program code and for calling tools, holds 256,000 tokens of context and can read images. At lowcloud it is processed on servers inside the EU.
Does my data go to China if I use Kimi K2.7 Code?
No. The request goes through lowcloud to Melious AI GmbH in Saarbrücken and is processed on servers inside the EU. Moonshot AI only publishes the openly available model weights and never sees your requests. Which European operator actually serves the model is not published by Melious, and we do not know it either.
Is Kimi K2.7 Code open source?
The weights are public, but the license is not plain MIT. Moonshot calls it a modified MIT license. It matches MIT except for one condition: above 100 million monthly active users or 20 million US dollars in monthly revenue, the name Kimi K2.7 Code has to appear in the user interface. For use inside lowcloud this has no effect.
Can Kimi K2.7 Code read images?
Yes. It evaluates screenshots, scans and photos. We tested that with images. Moonshot additionally lists video as experimental and only available through its own interface, so do not rely on video inside lowcloud.
Why does Kimi K2.7 Code reject fixed JSON schemas?
The vendor does not accept requests carrying a fixed JSON schema and answers with an error. We tested this against the interface ourselves. lowcloud then asks for the structure in the prompt. If a step's output must follow a specification exactly, use Kimi K2.6, GLM 5.1, GLM 5, Qwen 3.5 397B or DeepSeek V4 Flash.
Is there a newer model from Moonshot AI?
Yes, Kimi K3 from 16 July 2026, with roughly 1 million tokens of context. Kimi K2.7 Code is not discontinued because of it; Moonshot still lists it as the model for code and tool work. Both are in the lowcloud EU catalog.
Kimi K2.7 Code is a product name of Moonshot AI. There is no business relationship between lowcloud and Moonshot AI; we run the openly available model through a European provider. Last updated: 7 September 2026.